From Italy to Spain to improve weed detection through synthetic data

From June to September 2026, Leonardo Zecchin from Miningful carried out a secondment at the Centre for Automation and Robotics (CAR) of the Spanish National Research Council (CSIC), working in collaboration with Dionisio Andujar, Hugo Moreno, and Diego Torricelli, as part of the NeutraWeed project.

The secondment focused on the use of synthetic data to improve automatic weed detection in agricultural fields, contributing to the broader goal of NeutraWeed: supporting more selective weed management by enabling the reliable identification of plant species and distinguishing weeds that require intervention from plants that can be left undisturbed.

The research activities focused on generating synthetic agricultural images for training weed detection models. A diffusion-based generative model was fine-tuned using images from the training fields and used to generate synthetic weed instances, which were integrated into agricultural background images together with their corresponding object detection labels.

Different synthetic image datasets were created to investigate how visual realism, data diversity, and botanical correctness influence detection performance.

The experiments showed that including appropriately designed synthetic imagery can improve weed detection performance on previously unseen fields. The results also indicated that increasing visual realism alone does not necessarily improve detection, while data diversity and semantic correctness may play a more relevant role. During the secondment, synthetic image data were generated to support further investigation into the use of synthetic data for agricultural perception and automated weed management.

During the secondment, twelve synthetic image pools were generated, providing material for further investigation into the use of synthetic data for agricultural perception and automated weed management.

Finally, and importantly, the secondment strengthened the collaboration between Miningful and CAR-CSIC and supported further work on improving weed detection for more targeted weed management within NeutraWeed.